Fritter v1.0: Parakeet TDT 0.6B v2 for English dictation, no code-switching

A personal speech-recognition fine-tune for English dictation without code-switching: tech vocabulary (model and product names, CLI and cloud terms), brand names, acronyms, numbers and currency in digits, Irish-English phrasing, fast and slow speech, whispering, and everyday email and message dictation, all in one speaker's voice.

Name. Fritter is the code name of this model line; later retrains keep it and bump the version. The training run id is r11-general.

A full fine-tune of nvidia/parakeet-tdt-0.6b-v2 on one speaker's read-aloud phone recordings (an Irish English speaker living in Israel), run r11-general, trained 2026-10-27. Output keeps v2's casing and punctuation; numbers are written in digits.

Try it in your browser: danielrosehill/Fritter-Demo runs the int8 ONNX build on your own device.

Licence and access. Public, CC-BY-NC-4.0 (non-commercial). A personal project by Daniel Rosehill; not affiliated with NVIDIA.

Results

All models scored on the same test clips (dataset revision 520e7506e379, blend general test sets; never trained on, never used for checkpoint selection). WER is normalised (lowercase, no punctuation).

Model	Hebrew-free English WER (319 clips)	tech WER	own English WER (364)	LibriSpeech WER (200)	code-switching WER (614)	held-out Hebrew term recall	long-form WER
stock nvidia/parakeet-tdt-0.6b-v2	7.46%	7.53%	7.23%	1.23%	12.64%	24.90%	11.05%
this model (r11-general)	4.73%	4.01%	4.85%	1.47%	11.69%	24.27%	12.16%

On its target, Fritter cuts tech-sentence WER from 7.53% to 4.01% and Hebrew-free English WER from 7.46% to 4.73%, and recalls 79.6% of the tech and English target terms in the test takes against stock's 65.1% (447 occurrences of 262 terms; exact-case 74.1% vs 58.2%). Numbers recall rises from 33.9% to 50.0% (62 tokens). LibriSpeech is 0.24 points worse than stock (1.47% vs 1.23%, 200 utterances) and long-form dictation 1.1 points worse (12.16% vs 11.05%, 23 files), so general English outside the speaker's own voice is about where stock is, not better. It is not a code-switching model: its code-switching WER is only slightly better than stock's. Test takes are one speaker's, and their target terms are disjoint from training. Validation WER at the best epoch (17 of 20) was 4.24%.

Training data

Source: danielrosehill/Read-Aloud-Voice-Sentences at revision 520e7506e379; the mix is rebuilt from it by prepare (blend general).

• The speaker's takes that are not code-switching (is_code_switching() false). Every task subset is in except name-cipher and the held-back tags (weird voices); 10% LibriSpeech-PC replay.
• Mix: 319.9 min of the speaker's takes plus 35.6 min of LibriSpeech-PC replay (cased, punctuated; CC-BY-4.0) = 391.0 min, 3,848 clips.
• Validation: 1 in 16 of the speaker's train takes by id hash, used only to pick the checkpoint.

Training

NeMo 2.7.3 on one L40S (Modal), bf16. AdamW, lr 3e-05 (cosine, 10% warmup), 20 epochs, batch 16, seed 7. Best checkpoint best-epoch=17-val_wer=0.0424.ckpt, 38.2 min of training.

Usage

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download
path = hf_hub_download("danielrosehill/Fritter", "Fritter.nemo")
model = nemo_asr.models.ASRModel.restore_from(path)
print(model.transcribe(["clip.wav"])[0].text)   # 16 kHz mono

GGUF builds for transcribe.cpp (Offline Voice Input on Android, Handy on Linux) are in gguf/: Q8_0 (recommended), Q4_K_M and F16.

Limitations

• One speaker. It is adapted to one voice, accent and phone; others will see less benefit.
• The test sets are small (Hebrew-free English is 319 clips), so differences under about one WER point are within noise.
• Not trained on English-Hebrew code-switching: romanized Hebrew words come out as stock v2 writes them.
• Disfluent takes were trained on with clean targets, so the model may drop some fillers and repeated words.

Licence

CC-BY-NC-4.0. The base model, nvidia/parakeet-tdt-0.6b-v2, is CC-BY-4.0, by NVIDIA.